Asymmetric source-free unsupervised domain adaptation for medical image diagnosis

Abstract

Existing source-free unsupervised domain adaptation (SFUDA) methods primarily focus on addressing the domain gap issue for single-modal data, overlooking two crucial aspects: 1) In medical scenarios, clinicians often rely on multi-modal information for disease diagnosis. Consequently, emphasizing single-modal (symmetric modality) SFUDA algorithms neglect the complementary information from other modalities (asymmetric modalities). 2) Restricting SFUDA to a single modality limits downstream institutions’s ability to handle diverse modalities beyond that singular modality. To tackle these challenges, we propose an Asymmetric Source-Free Unsupervised Domain Adaptation (A-SFUDA) algorithm. This method leverages source model and unlabeled data from both symmetric and asymmetric modalities in the target domain for disease diagnosis. A-SFUDA adopts a two-stage training approach. In the first stage, A-SFUDA employs knowledge distillation (KD) to obtain two models capable of handling symmetric and asymmetric data in the target domain, facilitating preliminary diagnosis ability. In the second stage, A-SFUDA optimizes the target models through a pseudo-label correction mechanism based on multi-modal prediction correction and class-centered distance correction. Incorporating the two pseudo-label correction modules effectively mitigates noise within the training data, thereby facilitating the learning of the target models. We validate the performance of the proposed A-SFUDA algorithm on a large chest X-ray dataset, demonstrating its excellent performance for disease diagnosis in the target domain.